include/linux/huge_mm.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef _LINUX_HUGE_MM_H
   3 #define _LINUX_HUGE_MM_H
   4
   5 #include <linux/sched/coredump.h>
   6 #include <linux/mm_types.h>
   7
   8 #include <linux/fs.h> /* only for vma_is_dax() */
   9
  10 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf);
  11 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
  12                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
  13                   struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma);
  14 void huge_pmd_set_accessed(struct vm_fault *vmf);
  15 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
  16                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
  17                   struct vm_area_struct *vma);
  18
  19 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
  20 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud);
  21 #else
  22 static inline void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
  23 {
  24 }
  25 #endif
  26
  27 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf);
  28 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
  29                            pmd_t *pmd, unsigned long addr, unsigned long next);
  30 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, pmd_t *pmd,
  31                  unsigned long addr);
  32 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, pud_t *pud,
  33                  unsigned long addr);
  34 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
  35                    unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd);
  36 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
  37                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
  38                     unsigned long cp_flags);
  39
  40 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write);
  41 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write);
  42
  43 enum transparent_hugepage_flag {
  44         TRANSPARENT_HUGEPAGE_UNSUPPORTED,
  45         TRANSPARENT_HUGEPAGE_FLAG,
  46         TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
  47         TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
  48         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
  49         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
  50         TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
  51         TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG,
  52         TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG,
  53 };
  54
  55 struct kobject;
  56 struct kobj_attribute;
  57
  58 ssize_t single_hugepage_flag_store(struct kobject *kobj,
  59                                    struct kobj_attribute *attr,
  60                                    const char *buf, size_t count,
  61                                    enum transparent_hugepage_flag flag);
  62 ssize_t single_hugepage_flag_show(struct kobject *kobj,
  63                                   struct kobj_attribute *attr, char *buf,
  64                                   enum transparent_hugepage_flag flag);
  65 extern struct kobj_attribute shmem_enabled_attr;
  66
  67 #define HPAGE_PMD_ORDER (HPAGE_PMD_SHIFT-PAGE_SHIFT)
  68 #define HPAGE_PMD_NR (1<<HPAGE_PMD_ORDER)
  69
  70 /*
  71  * Mask of all large folio orders supported for anonymous THP; all orders up to
  72  * and including PMD_ORDER, except order-0 (which is not "huge") and order-1
  73  * (which is a limitation of the THP implementation).
  74  */
  75 #define THP_ORDERS_ALL_ANON     ((BIT(PMD_ORDER + 1) - 1) & ~(BIT(0) | BIT(1)))
  76
  77 /*
  78  * Mask of all large folio orders supported for file THP.
  79  */
  80 #define THP_ORDERS_ALL_FILE     (BIT(PMD_ORDER) | BIT(PUD_ORDER))
  81
  82 /*
  83  * Mask of all large folio orders supported for THP.
  84  */
  85 #define THP_ORDERS_ALL          (THP_ORDERS_ALL_ANON | THP_ORDERS_ALL_FILE)
  86
  87 #define thp_vma_allowable_order(vma, vm_flags, smaps, in_pf, enforce_sysfs, order) \
  88         (!!thp_vma_allowable_orders(vma, vm_flags, smaps, in_pf, enforce_sysfs, BIT(order)))
  89
  90 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
  91 #define HPAGE_PMD_SHIFT PMD_SHIFT
  92 #define HPAGE_PMD_SIZE  ((1UL) << HPAGE_PMD_SHIFT)
  93 #define HPAGE_PMD_MASK  (~(HPAGE_PMD_SIZE - 1))
  94
  95 #define HPAGE_PUD_SHIFT PUD_SHIFT
  96 #define HPAGE_PUD_SIZE  ((1UL) << HPAGE_PUD_SHIFT)
  97 #define HPAGE_PUD_MASK  (~(HPAGE_PUD_SIZE - 1))
  98
  99 extern unsigned long transparent_hugepage_flags;
 100 extern unsigned long huge_anon_orders_always;
 101 extern unsigned long huge_anon_orders_madvise;
 102 extern unsigned long huge_anon_orders_inherit;
 103
 104 static inline bool hugepage_global_enabled(void)
 105 {
 106         return transparent_hugepage_flags &
 107                         ((1<<TRANSPARENT_HUGEPAGE_FLAG) |
 108                         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG));
 109 }
 110
 111 static inline bool hugepage_global_always(void)
 112 {
 113         return transparent_hugepage_flags &
 114                         (1<<TRANSPARENT_HUGEPAGE_FLAG);
 115 }
 116
 117 static inline bool hugepage_flags_enabled(void)
 118 {
 119         /*
 120          * We cover both the anon and the file-backed case here; we must return
 121          * true if globally enabled, even when all anon sizes are set to never.
 122          * So we don't need to look at huge_anon_orders_inherit.
 123          */
 124         return hugepage_global_enabled() ||
 125                huge_anon_orders_always ||
 126                huge_anon_orders_madvise;
 127 }
 128
 129 static inline int highest_order(unsigned long orders)
 130 {
 131         return fls_long(orders) - 1;
 132 }
 133
 134 static inline int next_order(unsigned long *orders, int prev)
 135 {
 136         *orders &= ~BIT(prev);
 137         return highest_order(*orders);
 138 }
 139
 140 /*
 141  * Do the below checks:
 142  *   - For file vma, check if the linear page offset of vma is
 143  *     order-aligned within the file.  The hugepage is
 144  *     guaranteed to be order-aligned within the file, but we must
 145  *     check that the order-aligned addresses in the VMA map to
 146  *     order-aligned offsets within the file, else the hugepage will
 147  *     not be mappable.
 148  *   - For all vmas, check if the haddr is in an aligned hugepage
 149  *     area.
 150  */
 151 static inline bool thp_vma_suitable_order(struct vm_area_struct *vma,
 152                 unsigned long addr, int order)
 153 {
 154         unsigned long hpage_size = PAGE_SIZE << order;
 155         unsigned long haddr;
 156
 157         /* Don't have to check pgoff for anonymous vma */
 158         if (!vma_is_anonymous(vma)) {
 159                 if (!IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 160                                 hpage_size >> PAGE_SHIFT))
 161                         return false;
 162         }
 163
 164         haddr = ALIGN_DOWN(addr, hpage_size);
 165
 166         if (haddr < vma->vm_start || haddr + hpage_size > vma->vm_end)
 167                 return false;
 168         return true;
 169 }
 170
 171 /*
 172  * Filter the bitfield of input orders to the ones suitable for use in the vma.
 173  * See thp_vma_suitable_order().
 174  * All orders that pass the checks are returned as a bitfield.
 175  */
 176 static inline unsigned long thp_vma_suitable_orders(struct vm_area_struct *vma,
 177                 unsigned long addr, unsigned long orders)
 178 {
 179         int order;
 180
 181         /*
 182          * Iterate over orders, highest to lowest, removing orders that don't
 183          * meet alignment requirements from the set. Exit loop at first order
 184          * that meets requirements, since all lower orders must also meet
 185          * requirements.
 186          */
 187
 188         order = highest_order(orders);
 189
 190         while (orders) {
 191                 if (thp_vma_suitable_order(vma, addr, order))
 192                         break;
 193                 order = next_order(&orders, order);
 194         }
 195
 196         return orders;
 197 }
 198
 199 static inline bool file_thp_enabled(struct vm_area_struct *vma)
 200 {
 201         struct inode *inode;
 202
 203         if (!vma->vm_file)
 204                 return false;
 205
 206         inode = vma->vm_file->f_inode;
 207
 208         return (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS)) &&
 209                !inode_is_open_for_write(inode) && S_ISREG(inode->i_mode);
 210 }
 211
 212 unsigned long __thp_vma_allowable_orders(struct vm_area_struct *vma,
 213                                          unsigned long vm_flags, bool smaps,
 214                                          bool in_pf, bool enforce_sysfs,
 215                                          unsigned long orders);
 216
 217 /**
 218  * thp_vma_allowable_orders - determine hugepage orders that are allowed for vma
 219  * @vma:  the vm area to check
 220  * @vm_flags: use these vm_flags instead of vma->vm_flags
 221  * @smaps: whether answer will be used for smaps file
 222  * @in_pf: whether answer will be used by page fault handler
 223  * @enforce_sysfs: whether sysfs config should be taken into account
 224  * @orders: bitfield of all orders to consider
 225  *
 226  * Calculates the intersection of the requested hugepage orders and the allowed
 227  * hugepage orders for the provided vma. Permitted orders are encoded as a set
 228  * bit at the corresponding bit position (bit-2 corresponds to order-2, bit-3
 229  * corresponds to order-3, etc). Order-0 is never considered a hugepage order.
 230  *
 231  * Return: bitfield of orders allowed for hugepage in the vma. 0 if no hugepage
 232  * orders are allowed.
 233  */
 234 static inline
 235 unsigned long thp_vma_allowable_orders(struct vm_area_struct *vma,
 236                                        unsigned long vm_flags, bool smaps,
 237                                        bool in_pf, bool enforce_sysfs,
 238                                        unsigned long orders)
 239 {
 240         /* Optimization to check if required orders are enabled early. */
 241         if (enforce_sysfs && vma_is_anonymous(vma)) {
 242                 unsigned long mask = READ_ONCE(huge_anon_orders_always);
 243
 244                 if (vm_flags & VM_HUGEPAGE)
 245                         mask |= READ_ONCE(huge_anon_orders_madvise);
 246                 if (hugepage_global_always() ||
 247                     ((vm_flags & VM_HUGEPAGE) && hugepage_global_enabled()))
 248                         mask |= READ_ONCE(huge_anon_orders_inherit);
 249
 250                 orders &= mask;
 251                 if (!orders)
 252                         return 0;
 253         }
 254
 255         return __thp_vma_allowable_orders(vma, vm_flags, smaps, in_pf,
 256                                           enforce_sysfs, orders);
 257 }
 258
 259 #define transparent_hugepage_use_zero_page()                            \
 260         (transparent_hugepage_flags &                                   \
 261          (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
 262
 263 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 264                 unsigned long len, unsigned long pgoff, unsigned long flags);
 265
 266 void folio_prep_large_rmappable(struct folio *folio);
 267 bool can_split_folio(struct folio *folio, int *pextra_pins);
 268 int split_huge_page_to_list(struct page *page, struct list_head *list);
 269 static inline int split_huge_page(struct page *page)
 270 {
 271         return split_huge_page_to_list(page, NULL);
 272 }
 273 void deferred_split_folio(struct folio *folio);
 274
 275 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 276                 unsigned long address, bool freeze, struct folio *folio);
 277
 278 #define split_huge_pmd(__vma, __pmd, __address)                         \
 279         do {                                                            \
 280                 pmd_t *____pmd = (__pmd);                               \
 281                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 282                                         || pmd_devmap(*____pmd))        \
 283                         __split_huge_pmd(__vma, __pmd, __address,       \
 284                                                 false, NULL);           \
 285         }  while (0)
 286
 287
 288 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 289                 bool freeze, struct folio *folio);
 290
 291 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 292                 unsigned long address);
 293
 294 #define split_huge_pud(__vma, __pud, __address)                         \
 295         do {                                                            \
 296                 pud_t *____pud = (__pud);                               \
 297                 if (pud_trans_huge(*____pud)                            \
 298                                         || pud_devmap(*____pud))        \
 299                         __split_huge_pud(__vma, __pud, __address);      \
 300         }  while (0)
 301
 302 int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
 303                      int advice);
 304 int madvise_collapse(struct vm_area_struct *vma,
 305                      struct vm_area_struct **prev,
 306                      unsigned long start, unsigned long end);
 307 void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
 308                            unsigned long end, long adjust_next);
 309 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
 310 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
 311
 312 static inline int is_swap_pmd(pmd_t pmd)
 313 {
 314         return !pmd_none(pmd) && !pmd_present(pmd);
 315 }
 316
 317 /* mmap_lock must be held on entry */
 318 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 319                 struct vm_area_struct *vma)
 320 {
 321         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 322                 return __pmd_trans_huge_lock(pmd, vma);
 323         else
 324                 return NULL;
 325 }
 326 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 327                 struct vm_area_struct *vma)
 328 {
 329         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 330                 return __pud_trans_huge_lock(pud, vma);
 331         else
 332                 return NULL;
 333 }
 334
 335 /**
 336  * folio_test_pmd_mappable - Can we map this folio with a PMD?
 337  * @folio: The folio to test
 338  */
 339 static inline bool folio_test_pmd_mappable(struct folio *folio)
 340 {
 341         return folio_order(folio) >= HPAGE_PMD_ORDER;
 342 }
 343
 344 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 345                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 346 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 347                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 348
 349 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf);
 350
 351 extern struct page *huge_zero_page;
 352 extern unsigned long huge_zero_pfn;
 353
 354 static inline bool is_huge_zero_page(struct page *page)
 355 {
 356         return READ_ONCE(huge_zero_page) == page;
 357 }
 358
 359 static inline bool is_huge_zero_pmd(pmd_t pmd)
 360 {
 361         return pmd_present(pmd) && READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd);
 362 }
 363
 364 static inline bool is_huge_zero_pud(pud_t pud)
 365 {
 366         return false;
 367 }
 368
 369 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 370 void mm_put_huge_zero_page(struct mm_struct *mm);
 371
 372 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 373
 374 static inline bool thp_migration_supported(void)
 375 {
 376         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 377 }
 378
 379 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 380 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 381 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 382 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 383
 384 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 385 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 386 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 387
 388 static inline bool folio_test_pmd_mappable(struct folio *folio)
 389 {
 390         return false;
 391 }
 392
 393 static inline bool thp_vma_suitable_order(struct vm_area_struct *vma,
 394                 unsigned long addr, int order)
 395 {
 396         return false;
 397 }
 398
 399 static inline unsigned long thp_vma_suitable_orders(struct vm_area_struct *vma,
 400                 unsigned long addr, unsigned long orders)
 401 {
 402         return 0;
 403 }
 404
 405 static inline unsigned long thp_vma_allowable_orders(struct vm_area_struct *vma,
 406                                         unsigned long vm_flags, bool smaps,
 407                                         bool in_pf, bool enforce_sysfs,
 408                                         unsigned long orders)
 409 {
 410         return 0;
 411 }
 412
 413 static inline void folio_prep_large_rmappable(struct folio *folio) {}
 414
 415 #define transparent_hugepage_flags 0UL
 416
 417 #define thp_get_unmapped_area   NULL
 418
 419 static inline bool
 420 can_split_folio(struct folio *folio, int *pextra_pins)
 421 {
 422         return false;
 423 }
 424 static inline int
 425 split_huge_page_to_list(struct page *page, struct list_head *list)
 426 {
 427         return 0;
 428 }
 429 static inline int split_huge_page(struct page *page)
 430 {
 431         return 0;
 432 }
 433 static inline void deferred_split_folio(struct folio *folio) {}
 434 #define split_huge_pmd(__vma, __pmd, __address) \
 435         do { } while (0)
 436
 437 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 438                 unsigned long address, bool freeze, struct folio *folio) {}
 439 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 440                 unsigned long address, bool freeze, struct folio *folio) {}
 441
 442 #define split_huge_pud(__vma, __pmd, __address) \
 443         do { } while (0)
 444
 445 static inline int hugepage_madvise(struct vm_area_struct *vma,
 446                                    unsigned long *vm_flags, int advice)
 447 {
 448         return -EINVAL;
 449 }
 450
 451 static inline int madvise_collapse(struct vm_area_struct *vma,
 452                                    struct vm_area_struct **prev,
 453                                    unsigned long start, unsigned long end)
 454 {
 455         return -EINVAL;
 456 }
 457
 458 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 459                                          unsigned long start,
 460                                          unsigned long end,
 461                                          long adjust_next)
 462 {
 463 }
 464 static inline int is_swap_pmd(pmd_t pmd)
 465 {
 466         return 0;
 467 }
 468 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 469                 struct vm_area_struct *vma)
 470 {
 471         return NULL;
 472 }
 473 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 474                 struct vm_area_struct *vma)
 475 {
 476         return NULL;
 477 }
 478
 479 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
 480 {
 481         return 0;
 482 }
 483
 484 static inline bool is_huge_zero_page(struct page *page)
 485 {
 486         return false;
 487 }
 488
 489 static inline bool is_huge_zero_pmd(pmd_t pmd)
 490 {
 491         return false;
 492 }
 493
 494 static inline bool is_huge_zero_pud(pud_t pud)
 495 {
 496         return false;
 497 }
 498
 499 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 500 {
 501         return;
 502 }
 503
 504 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 505         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 506 {
 507         return NULL;
 508 }
 509
 510 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 511         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 512 {
 513         return NULL;
 514 }
 515
 516 static inline bool thp_migration_supported(void)
 517 {
 518         return false;
 519 }
 520 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 521
 522 static inline int split_folio_to_list(struct folio *folio,
 523                 struct list_head *list)
 524 {
 525         return split_huge_page_to_list(&folio->page, list);
 526 }
 527
 528 static inline int split_folio(struct folio *folio)
 529 {
 530         return split_folio_to_list(folio, NULL);
 531 }
 532
 533 /*
 534  * archs that select ARCH_WANTS_THP_SWAP but don't support THP_SWP due to
 535  * limitations in the implementation like arm64 MTE can override this to
 536  * false
 537  */
 538 #ifndef arch_thp_swp_supported
 539 static inline bool arch_thp_swp_supported(void)
 540 {
 541         return true;
 542 }
 543 #endif
 544
 545 #endif /* _LINUX_HUGE_MM_H */